tomoyo: Don't check open/getattr permission on sockets.

Commit Message

Tetsuo Handa June 9, 2019, 6:41 a.m. UTC

syzbot is reporting that use of SOCKET_I()->sk from open() can result in
use after free problem [1], for socket's inode is still reachable via
/proc/pid/fd/n despite destruction of SOCKET_I()->sk already completed.

But there is no point with calling security_file_open() on sockets
because open("/proc/pid/fd/n", !O_PATH) on sockets fails with -ENXIO.

There is some point with calling security_inode_getattr() on sockets
because stat("/proc/pid/fd/n") and fstat(open("/proc/pid/fd/n", O_PATH))
are valid. If we want to access "struct sock"->sk_{family,type,protocol}
fields, we will need to use security_socket_post_create() hook and
security_inode_free() hook in order to remember these fields because
security_sk_free() hook is called before the inode is destructed. But
since information which can be protected by checking
security_inode_getattr() on sockets is trivial, let's not be bothered by
"struct inode"->i_security management.

There is point with calling security_file_ioctl() on sockets. Since
ioctl(open("/proc/pid/fd/n", O_PATH)) is invalid, security_file_ioctl()
on sockets should remain safe.

[1] https://syzkaller.appspot.com/bug?id=73d590010454403d55164cca23bd0565b1eb3b74

Signed-off-by: Tetsuo Handa <[e-mail]>
Reported-by: syzbot <[e-mail]>
---
 security/tomoyo/tomoyo.c | 7 +++++++
 1 file changed, 7 insertions(+)

Comments

Tetsuo Handa June 16, 2019, 6:49 a.m. UTC | #1

Hello, Al.

Q1: Do you agree that we should fix TOMOYO side rather than SOCKET_I()->sk
    management.

Q2: Do you see any problem with using f->f_path.dentry->d_inode ?
    Do we need to use d_backing_inode() or d_inode() ?

Regards.

On 2019/06/09 15:41, Tetsuo Handa wrote:
> syzbot is reporting that use of SOCKET_I()->sk from open() can result in
> use after free problem [1], for socket's inode is still reachable via
> /proc/pid/fd/n despite destruction of SOCKET_I()->sk already completed.
> 
> But there is no point with calling security_file_open() on sockets
> because open("/proc/pid/fd/n", !O_PATH) on sockets fails with -ENXIO.
> 
> There is some point with calling security_inode_getattr() on sockets
> because stat("/proc/pid/fd/n") and fstat(open("/proc/pid/fd/n", O_PATH))
> are valid. If we want to access "struct sock"->sk_{family,type,protocol}
> fields, we will need to use security_socket_post_create() hook and
> security_inode_free() hook in order to remember these fields because
> security_sk_free() hook is called before the inode is destructed. But
> since information which can be protected by checking
> security_inode_getattr() on sockets is trivial, let's not be bothered by
> "struct inode"->i_security management.
> 
> There is point with calling security_file_ioctl() on sockets. Since
> ioctl(open("/proc/pid/fd/n", O_PATH)) is invalid, security_file_ioctl()
> on sockets should remain safe.
> 
> [1] https://syzkaller.appspot.com/bug?id=73d590010454403d55164cca23bd0565b1eb3b74
> 
> Signed-off-by: Tetsuo Handa <[e-mail]>
> Reported-by: syzbot <[e-mail]>
> ---
>  security/tomoyo/tomoyo.c | 7 +++++++
>  1 file changed, 7 insertions(+)
> 
> diff --git a/security/tomoyo/tomoyo.c b/security/tomoyo/tomoyo.c
> index 716c92e..9661b86 100644
> --- a/security/tomoyo/tomoyo.c
> +++ b/security/tomoyo/tomoyo.c
> @@ -126,6 +126,9 @@ static int tomoyo_bprm_check_security(struct linux_binprm *bprm)
>   */
>  static int tomoyo_inode_getattr(const struct path *path)
>  {
> +	/* It is not safe to call tomoyo_get_socket_name(). */
> +	if (path->dentry->d_inode && S_ISSOCK(path->dentry->d_inode->i_mode))
> +		return 0;
>  	return tomoyo_path_perm(TOMOYO_TYPE_GETATTR, path, NULL);
>  }
>  
> @@ -316,6 +319,10 @@ static int tomoyo_file_open(struct file *f)
>  	/* Don't check read permission here if called from do_execve(). */
>  	if (current->in_execve)
>  		return 0;
> +	/* Sockets can't be opened by open(). */
> +	if (f->f_path.dentry->d_inode &&
> +	    S_ISSOCK(f->f_path.dentry->d_inode->i_mode))
> +		return 0;
>  	return tomoyo_check_open_permission(tomoyo_domain(), &f->f_path,
>  					    f->f_flags);
>  }
>

Al Viro June 18, 2019, 8:49 p.m. UTC | #2

On Sun, Jun 16, 2019 at 03:49:00PM +0900, Tetsuo Handa wrote:
> Hello, Al.
> 
> Q1: Do you agree that we should fix TOMOYO side rather than SOCKET_I()->sk
>     management.

You do realize that sockets are not unique in that respect, right?
All kinds of interesting stuff can be accessed via /proc/*/fd/*, and
it _can_ be closed under you.  So I'd suggest checking how your code
copes with similar for pipes, FIFOs, epoll, etc., accessed that way...

We are _not_ going to be checking that in fs/open.c - the stuff found
via /proc/*/fd/* can have the associated file closed by the time
we get to calling ->open() and we won't know that until said call.

> Q2: Do you see any problem with using f->f_path.dentry->d_inode ?
>     Do we need to use d_backing_inode() or d_inode() ?

Huh?  What's wrong with file_inode(f), in the first place?  And
just when can that be NULL, while we are at it?

> >  static int tomoyo_inode_getattr(const struct path *path)
> >  {
> > +	/* It is not safe to call tomoyo_get_socket_name(). */
> > +	if (path->dentry->d_inode && S_ISSOCK(path->dentry->d_inode->i_mode))
> > +		return 0;

Can that be called for a negative?

Patch

diff --git a/security/tomoyo/tomoyo.c b/security/tomoyo/tomoyo.c
index 716c92e..9661b86 100644
--- a/security/tomoyo/tomoyo.c
+++ b/security/tomoyo/tomoyo.c
@@ -126,6 +126,9 @@  static int tomoyo_bprm_check_security(struct linux_binprm *bprm)
  */
 static int tomoyo_inode_getattr(const struct path *path)
 {
+	/* It is not safe to call tomoyo_get_socket_name(). */
+	if (path->dentry->d_inode && S_ISSOCK(path->dentry->d_inode->i_mode))
+		return 0;
 	return tomoyo_path_perm(TOMOYO_TYPE_GETATTR, path, NULL);
 }
 
@@ -316,6 +319,10 @@  static int tomoyo_file_open(struct file *f)
 	/* Don't check read permission here if called from do_execve(). */
 	if (current->in_execve)
 		return 0;
+	/* Sockets can't be opened by open(). */
+	if (f->f_path.dentry->d_inode &&
+	    S_ISSOCK(f->f_path.dentry->d_inode->i_mode))
+		return 0;
 	return tomoyo_check_open_permission(tomoyo_domain(), &f->f_path,
 					    f->f_flags);
 }